Staff time is one level of evidence
The GDS cross-government trial reported 26 minutes saved per user per day in office work. It did not measure an equivalent reduction in service queues.
UK Government Digital Service04 / GOVERNMENT
Faster tasks.
Better services?
Investigating when AI saves time across a whole public service, and when the delay simply moves to another desk.
01 / The question
Does the whole
journey improve?
An assistant can make drafting faster while applicants wait just as long. Completion depends on connected teams, evidence, review capacity, and demand. We want to measure the service people receive, including corrections and the work they must do themselves.
Measurement changes the claim
HMRC’s July 2026 evaluation reports adjusted, self-reported time savings and a small task exercise. These help frame a study; they are not a direct measure of applicant outcomes.
HMRC02 / Make the argument testable
The slowest step
sets the ceiling.
In our fictional service, intake can prepare 120 cases daily but review can process only 80. Doubling intake capacity leaves completion at 80. If more cases need a second review, the completion ceiling falls.
- 01 / INTAKE120 → 240
Cases prepared per day.
- 02 / REVIEW80
Review passes per day.
THE BOTTLENECK - 03 / FINAL PROCESSING100
Reviewed cases closed per day.
One review per case80 completions / day
20% need one extra review66.7 completions / day
Invented capacities, not a government benchmark. Assumes an existing backlog, enough incoming cases, and equal effort per review pass. The rework scenario assumes exactly one additional pass for 20% of cases: 80 ÷ 1.2 ≈ 66.7.
03 / The investigation we would run
Follow a case
from start to resolution.
Map one service before introducing a tool. Separate active work from waiting, then test assistance against historical or synthetic cases before considering a live evaluation.
PROPOSED STUDY
A research plan informed by published evidence. Data collection has not begun.
- 01
Correct resolution
Time to an answer that survives an agreed quality review, with unresolved cases still visible.
- 02
Repeat work
Returned cases, corrections, duplicate contacts, and the reason each return happened.
- 03
Applicant effort
Time spent obtaining evidence, asking for help, and restarting an incomplete application.
- 04
Distribution of waits
Long waits and complex cases alongside the median. Track who drops out or cannot get access.
What would change our view?
If most delay comes from unavailable external records, improving the connection may be more useful than drafting assistance. If faster intake increases rework, fix evidence capture before expanding the tool. Staff-time savings can still matter without being called faster public services.
04 / The full argument
ANALYSIS / GOVERNMENT / 7 MIN READ
Faster paperwork. Same damn queue.
The assistant writes in seconds. The application still takes weeks. What happens between those two facts is the part worth fixing.
READ THE ARTICLE↗ / Where this could lead
A question worth working on.
We want to help agencies improve dependable service capacity while preserving accountable decisions and a practical route for people to question them.